Assessing Personalized Software Defect Predictors

Eken B.

40th ACM/IEEE International Conference on Software Engineering (ICSE), Gothenburg, Sweden, 27 May - 03 June 2018, pp.488-491 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1145/3183440.3183449
  • City: Gothenburg
  • Country: Sweden
  • Page Numbers: pp.488-491
  • Istanbul Technical University Affiliated: Yes


Software defect prediction models guide developers and testers to identify defect prone software modules in fewer time and effort, compared to manual inspections of the source code. The state-of-the-art predictors on publicly available software engineering data could catch around 70% of the defects. While early studies mostly utilize static code properties of the software, recent studies incorporate the people factor into the prediction models, such as the number of developers that touched a code unit, the experience of the developer, and interaction and cognitive behaviors of developers. Those information could give a stronger clue about the defect-prone parts because they could explain defect injection patterns in software development. Personalization has been emerging in many other systems such as social platforms, web search engines such that people get customized recommendations based on their actions, profiles and interest. Following this point of view, customization in defect prediction with respect to each developer would increase predictions' accuracy and usefulness than traditional, general models. In this thesis, we focus on building a personalized defect prediction framework that gives instant feedback to the developer at change level, based on historical defect and change data. Our preliminary analysis of the personalized prediction models of 121 developers in six open source projects indicate that, a personalized approach is not always the best model when compared to general models built for six projects. Other factors such as project characteristics, developer's historical data, the context and frequency of contributions, and/or development methodologies might affect which model to consider in practice. Eventually, this topic is open to improvement with further empirical studies on each of these factors.